outside the classroom: evidence on non-instructional ... · effects on student learning. in fact, a...
TRANSCRIPT
1
Outside the Classroom:
Evidence on Non-Instructional Spending and Student Outcomes*
Lucy Sorensen
Sanford School of Public Policy, Duke University
Draft Prepared for Annual AEFP 41st Annual Conference
Denver, CO, March 17-19 *Work on this paper was supported by a pre-doctoral fellowship provided by the National Institute of Child Health and Development (T32-HD0736) through the Center for Developmental Science, University of North Carolina at Chapel Hill. The author is grateful for data assistance from the North Carolina Education Research Data Center (NCERDC) and the North Carolina Department of Public Instruction (NC DPI). The author would also like to thank Kenneth Dodge, Helen Ladd, Philip Cook, Marcos Rangel, and Joseph Hotz, for invaluable feedback.
2
Abstract
Much prior research on school finance has assessed the relation between overall
funding of schools and student achievement outcomes. This study moves beyond the
simple bivariate association to contribute new evidence on how non-instructional
investments – such as increased spending on school social workers, guidance counselors,
and health services – affect multiple aspects of student performance and well-being.
Merging several administrative data sources spanning the 1996-2013 school years in
North Carolina, I use an instrumental variables approach to estimate the extent to which
local expenditure shifts affect students’ academic and behavioral outcomes. My findings
indicate that exogenous increases in spending on non-instructional services not only
reduce student absenteeism and disciplinary problems (important predictors of long-term
outcomes) but also significantly raise student achievement, in similar magnitude to
corresponding increases in instructional spending. Furthermore, subgroup analyses
suggest that investments in student support personnel such as social workers, health
services, and guidance counselors, in schools with concentrated low-income student
populations could go a long way toward closing socioeconomic achievement gaps.
3
1. Introduction
A number of studies, with varying results, have examined the link between school
resources and student achievement (summarized in Greenwald, Hedges, & Laine, 1996;
Hanushek, 1997; and Verstegen & King, 1998). The central debate in this school finance
literature concerns whether increased education spending translates to higher student test
scores. This body of research, however, largely fails to consider how resources are
allocated across different functions within schools, or how increased spending may affect
students across a broader set of student outcome domains. The current study seeks to
address both of those topics, with particular attention to whether increased expenditures
on social workers, health services, and guidance services within schools contribute to
aspects of student well-being.
For the first time, in 2012 more than half of all U.S. students in K-12 schools
came from low-income families, defined as meeting the eligibility requirements for free
or reduced-price school lunches (NCES, 2013). Employers lament American
adolescents’ low high school graduation rates (despite recent improvements) and even
lower readiness to join the work force. At the same time, the rate of children and youth
with diagnosable mental health disorders in the U.S. has grown to over 20 percent, and
other health concerns for students such as obesity, allergies, and asthma continue to rise
(CDC, 2013). School systems across the country are considering the extent to which they
should provide support services for the types of challenges that students bring to the
classroom from outside of school, with little research-based evidence to help guide
policymakers’ decisions.
4
Although salaries and benefits for teachers account for over half of public school
expenditures, spending on support services for students such as health, guidance, and
social work services has grown in prominence. Between the 1990-91 and 2010-11 school
years, inflation-adjusted per-pupil expenditures nationwide on student support services
increased by 66 percent (NCES, 2013, Table 236.60). The recession of 2008 altered the
upward trajectory of spending, and local districts responded in different ways by
curtailing spending for different purposes. The question of how this temporal and
geographic variation in the mix of type of education spending affects student academic
and behavioral outcomes has gone largely unanswered.
The current study focuses on the case of North Carolina, and involves two main
components. First, I examine descriptive trends in instructional and non-instructional
spending over the past 20 years. Second, I assess whether shifts in the levels of local
instructional and non-instructional spending affect students’ academic and behavioral
outcomes. The study relies on administrative data merged from several sources, spanning
the 1995-1996 through 2012-2013 school years.1 County-level education expenditures
by detailed purpose code and year come from the North Carolina Department of Public
Instruction (DPI). Student-level data on standardized test scores, absences, disciplinary
offenses, and extracurricular participation come from the North Carolina Education
Research Data Center (NCERDC).
In North Carolina, as in all other states, school funding derives from local
(26.2%), state (62.1%), and federal (11.7%) revenue sources. The state funds schools
with three basic types of allotments: position allotments, for which a certain number of
certified teachers and other educators are provided to local school systems and paid based 1 From this point forward, I will refer to this time period as from 1996-2013 for simplicity.
5
on the state salary schedule; dollar allotments, which can be used to hire employees or
purchase goods for a specific purpose; and categorical allotments, which provide as much
funding as necessary to address the needs of a particular population or service. County
governments in North Carolina are responsible for fulfilling school facility requirements
and also for supplementing current operating expenses of the public school system
“within the financial resources and consistent with the fiscal policies of the board of
county commissioners” (North Carolina General Statutes, 2014, 115C-408). For this
study, I focus primarily on the impacts of local education expenditures based on two
considerations. First, county governments tend to have more flexibility in terms of both
their level of educational spending and also to which school resources they allocate
funds. Second, my identification strategy relies on quasi-random changes to local
revenues and governing bodies, which have no direct relation to state or federal spending.
This analysis uses an instrumental variable approach, taking advantage of
exogenous variation in county expenditures on different educational categories. This
variation in local education spending comes from two different sources. First, every
eight years (at a minimum) each North Carolina county must reassess property values.
When counties do so, they tend to experience sudden and sizeable increases in property
tax revenues, and a disproportionate amount goes to abrupt increases in local school
budgets. Because revaluation years differ across counties on a schedule that is
uncorrelated with county factors, I can use these within-county discontinuities at
revaluation year as quasi-random variation in local expenditures. Second, since boards of
county commissioners have direct control over both the level and types of funding for
education in North Carolina, changes in the political party majority of boards affect both
6
total county expenditures and the amounts of resources allocated toward academic and
non-instructional educational purposes. This within-county measure of political change
therefore can serve as a second instrumental variable. (An interaction term between the
revaluation variable and the political variable serves as a third).
This analysis examines for the first time how shifts in absolute and proportional
school financing for non-instructional support services at the district level might
influence students’ development. It estimates causal effects of student support-services
expenditures on achievement and several “non-cognitive” outcomes (including
absenteeism, disciplinary infractions, and extracurricular participation) shown to be
predictive of long-term well-being (Jackson, 2012). Recent empirical research has
further emphasized the long-term importance of non-academic behaviors for educational
attainment, labor market outcomes, and social outcomes (Heckman, Stixrud, & Urzua,
2006). In estimating effects on both student achievement and behavioral outcomes, the
analysis uses a two-stage least squares approach alluded to above to address potential
empirical challenges: first, that district decision-making about budgetary allocations is
not random; and second, that there may be a time lag between current expenditures and
effects on student outcomes. (Specifically, to account for multiple endogenous predictor
variables, I use full information generalized two stage least squares estimates of a system
of simultaneous equations). The estimation models also control for concurrent spending
on other purposes, county and year fixed effects, detailed student-level covariates, and
county time trends.
Results show that increased spending on non-instructional services – including
social workers, guidance counselors, and health services – not only reduces student
7
absenteeism and disciplinary problems, but also significantly increases student
achievement across all grades and subjects. In fact, the academic benefits of increasing
non-instructional resources at the county level match those of increasing instructional
salaries by the same amount. My findings indicate that spending only $100 extra per
pupil on non-instructional services could increase student test scores across various
subjects by between 0.014 and 0.099 standard deviations. By comparison, spending $100
extra per pupil on local salary supplements for teachers would increase average student
achievement by 0.016 to 0.078 standard deviations.2 Non-instructional spending
significantly outperforms instructional spending for math results, but the opposite is true
for reading.
These non-instructional effects on learning are substantially larger in schools with
high concentrations of poverty compared to those with lower concentrations of poverty,
with effect sizes in high-poverty schools up to 0.239 standard deviations (roughly
equivalent to average test score growth over the course of a school year) for an additional
$100 per pupil. This pattern suggests that investing in school support services and
opportunities for outside-of-school enrichment may hold greater value for students in
low-income communities. Such a robust effect for students in high-poverty schools
indicates that health and social services provision could be one low-cost policy option for
stemming the tide of widening socioeconomic achievement gaps (Reardon, 2011).
A potential mechanism is suggested by the fact that increased non-instructional
spending also significantly reduces student absences and serious disciplinary infractions.
$100 per pupil toward school support services translates to 0.58 fewer absences per 2 These values are presented as ranges of test score increases across four groups: math scores in grades 4-8, reading scores in grades 4-8, humanities scores in grades 9-12, and science scores in grades 9-12. Detailed findings by student subgroup are presented in the results section.
8
student per year and 0.99 fewer serious disciplinary infractions. If social workers,
truancy officers, or health services, for example, are capable of improving student
attendance and behavior indicators, then the mere fact that students are at school more
frequently, with fewer behavioral disruptions, could drive a large portion of the ultimate
effects on student learning. In fact, a simple mediation analysis reveals that reductions in
absences and disciplinary infractions explain up to 30 percent of the main effect of
instructional expenditures on student test scores and up to 88 percent of the main effect of
non-instructional expenditures.
Section 2 provides contextual information on prior research regarding both school
resources and the school finance landscape in North Carolina. Section 3 introduces a
simple model of local government decision-making for optimizing educational outcomes.
Section 4 describes the new dataset used in this study and the empirical approach for
estimating effects of non-instructional expenditures. Section 5 presents results. Section 6
discusses the relevance of these findings for education finance at the local, state, and
federal levels.
2. Background
Prior research on non-instructional school resources
The literature on returns to school resources has a long history, with a back-and-
forth debate centered on whether or not money matters. This often equates to whether
class size matters. Hanushek (1997) concludes from a meta-analysis that variation in
school resources within the typical range across districts is not consistently related to
student achievement, although Greenwald, Hedges, and Laine (1996) perform a similar
9
meta-analysis with findings opposite to Hanushek’s primary conclusions. Many of the
earlier studies face empirical problems because cross-sectional studies generate biased
estimates of returns to school spending, though the direction of this bias is debated
(Haegeland, Raaum, & Salvanes, 2012).
I assert that the domains to which resources are allocated are as important as the
overall level of resources. This topic – impacts of non-instructional student services and
opportunities – has a much more limited evidence base, however. Dee (2005) finds at the
national level that instructional spending increases high school graduation rates, whereas
other types of spending have a small negative effect on graduation rates. Reback (2010a)
explores the effects of school counselor policies on student achievement and behavioral
problems. He finds that greater counselor subsidies reduce the frequency of disciplinary
incidents but do not strongly influence achievement. In another study (Reback, 2010b),
he finds that elementary counselors substantially influence teachers’ perceptions of
school climate, reduce the fraction of teachers reporting that their instruction suffers due
to student misbehavior, and reduce the fraction reporting problems with students
physically fighting each other, cutting class, stealing, or using drugs. Carrell and
Hoekstra (2011) conclude that one additional school counselor reduces student
misbehavior and increases boys’ academic achievement by over one percentile point.
These three papers resemble the motivation of the current study, but I use a different
identification strategy and consider other elements of non-instructional resources such as
school social workers and health services.
Researchers and education leaders increasingly recognize the potential value of
health services in schools, including through implementation of school-based health
10
centers (SBHCs). Lovenheim, Reback, and Wedenoja (2015) examine the impacts of
providing primary care health services to students from low-income families (delivered
through SBHCs) and find large reductions in teen fertility, though no effects on high
school dropout rates. Although causal analysis on this topic is still thin, a sizeable
amount of work has demonstrated that poor health or adverse health events among
children are associated with worse long-run educational and economic outcomes (e.g.
Currie et al. 2010). Therefore, it logically follows that the provision of health services
within schools may prevent such adverse outcomes or have advantageous spillovers for
student learning.
The third main school resource of interest in this study, in addition to counselors
and health services, is school social work and attendance services. School social workers
serve as mental health providers for students and can assist in linking students and their
families to a variety of community services. There are currently no large-scale studies of
the effects of social work services in schools. However, prior research has evaluated a
variety of types of small-scale social work interventions. A meta-analysis of universal
and targeted social work interventions by Allen-Meares, Montgomery, and Kim (2013)
illustrates that many of these programs have positive effects in terms of student sexual
health, aggression, self-esteem, school attendance, identity, and depression.
A recent large-scale study by Jackson, Johnson, and Persico (2015) uses variation
in school spending from court-ordered school finance reforms to determine that increased
total per-pupil funding has large positive effects on students in high-poverty districts, but
smaller or nonexistent effects for students in more affluent districts. This finding could
reflect a number of factors. Relevant to the current study, though, it could support the
11
hypothesis that certain services provided within schools – such as social workers,
guidance counselors, psychologists, and health services – are more valuable to students
whose families lack the means to pay for these services and other opportunities outside of
school. The current study will test this hypothesis more formally.
North Carolina school finance
Understanding the fiscal context of the North Carolina public school system is
essential for examining the role that non-instructional expenditures play in fostering
positive student outcomes. As in all states, funding for education comes from a
combination of federal, state, and local sources. In North Carolina, local revenues
account for about 26 percent of total education spending and a greater share of school
capital spending and salaries for non-certified personnel (NC DPI, 2015). Although the
current study examines the effects of aggregate spending (local, federal, and state) on
student outcomes, there tends to be greater flexibility and geographic variation in how
local funds are appropriated, and the local share of current education expenses has been
growing over time (North Carolina Center for County Research, 2015). For this reason,
research such as the current study aimed toward evaluating local education finance
decision-making holds immediate policy relevance.
An elected board of county commissioners is responsible for determining the
yearly public education budget for a given county and for allocating local revenues
among various educational and non-educational purposes. A majority of county revenues
in the 2013-2014 year originated from property taxes (51%), with the remainder coming
from sales taxes, sales and services, intergovernmental transfers, debt services, and other
12
sources. The county commissioners allocate these local revenues toward a number of
different purposes: education (35%); public safety (20%); human services (20%); general
government (10%); and other categories (NCACC, 2015).3 In the domain of public
education, North Carolina counties are charged with building, equipping, and maintaining
school facilities and also with supplementing school operating expenses. They provide
current expense funds in a lump-sum annual appropriation to school districts. These
funds are typically used for teacher salary supplements, provision of non-certified
positions, and administrator positions. The fifteen city districts that exist within county
boundaries have no fiscal authority, and therefore are considered as a single unit together
with the rest of the county for this analysis. 4
Nationwide, 23.4 percent of operating expenditures nationwide in unified school
districts in 47 states are dedicated to non-instructional purposes (Dee, 2005). This
translates to around $935 per pupil for total non-instructional salary spending and $3,024
per pupil for total instructional salary spending. Larger, county-wide local education
agencies (LEAs) tend to allocate more money per student toward non-instructional
purposes than do smaller, community-centric LEAs (Deluca, 2015). In this study, I
define instructional spending as any direct salary expenditures on instructional services.
In North Carolina, the state pays teachers through a statewide standardized salary
schedule. All local instructional expenditures at the county level therefore are in the form
3 These percentages reflect the average across North Carolina’s 100 counties; counties vary in their spending patterns. 4 Technically, the reality is more complicated. Counties are responsible for raising funds and allocating those funds to districts within the county. Counties generally allocate funds to particular purposes and functions, but school boards can amend these numbers by up to 25%. There is a back-and-forth budgeting process between school boards and county governments in which LEAs hire school personnel, establish salary supplements, and determine school facility needs, and the county is tasked with providing “sufficient” financial resources.
13
of either salary supplements to the state-funded salary level or salaries for teaching
assistants.
I define non-instructional spending as any direct salary expenditures on 1)
guidance and psychological services; 2) social work and attendance services; and 3)
health services.5 Unlike with teachers, counties and local education agencies both have
the option of using local revenues to hire school personnel in these non-instructional
categories. These three expenditure categories have the practical advantage of being
measured consistently across the time period of interest (1996-2013), unlike other
purpose codes. They also reflect important aspects of students’ outside-the-classroom
development. Social workers address attendance issues and help students overcome
obstacles to learning that they may face in their homes or neighborhoods. Guidance
counselors often assist students in their personal and social development, and they
support students in building and achieving their long-term educational and career plans.
Health services can provide supplementary medical, dental, and nursing care.
Trends in spending in North Carolina
Education spending, in each of its various components, has grown and waned
during the past two decades in North Carolina. Both economic conditions and state and
federal political changes have driven much of this variation in school expenditures. In
Figure 1a, a binned scatter plot of county-year observations with a quadratic fit plot, we
observe that spending for instructional salaries rose between 1996 and the early 2000s,
but have experienced a steady decline since the 2004-2005 school year, brought on by 5 The exact purpose codes for these expenditures are: 5820 Attendance – social work services; 5830 Guidance services; and 5840 Health services. Exact descriptions of these categories are provided in Appendix Table A1.
14
severe teacher salary cuts from the North Carolina state legislature and other cutbacks
due to the economic recession. One can notice that there are no important differences
between instructional spending patterns of high-poverty counties and those of low-
poverty counties.
On the other hand, for non-instructional spending we observe a large differential
between expenditures in low-poverty and high-poverty counties starting around 2005
(See Figure 1b). Low-poverty counties, defined as those with a median poverty rate of
school-age individuals between 1996 and 2013 lower than 25 percent, exhibited linearly
increasing expenditures on non-instructional support services (attendance and social
work, guidance and psychological services, and health services) across the time period.6
High-poverty counties display the same general trend, but with much higher per-pupil
expenditures on non-instructional services since around 2005. One potential driving
force behind this change was implementation of the Child and Family Support Team
Initiative (CFST). Beginning in 2005-2006, state funds supported one certified school
nurse and one licensed school social worker in each of 101 schools across the state with a
large proportion of high-risk students (Gifford et al., 2010).
Maps in Appendix Figure A1 illustrate both the temporal and geographic
variation in expenditures across North Carolina’s one hundred counties. Some of this
variation reflects state allotment policies: counties designated as “low-wealth” or as
“small-county” receive additional per-pupil funds from the state. Other variation
represents factors such as differences in local revenues, differences in how local
governments choose to spend those revenues, and federal grant program receipts. These
6 All current expenditures are inflation-adjusted using the urban consumer price index (BLS 2015) to represent real 2013 dollars.
15
maps suggest a large degree of heterogeneity in how local entities allocate resources
among different educational purposes, and how the spending patterns change over time.
3. Model of county educational investments
A local education agency (LEA), and its corresponding county governing body,
have many overarching objectives and must make decisions and tradeoffs with these
multiple goals in mind. However, given the current emphasis on test-based
accountability, it is likely that the level of student proficiency on standardized tests is
high on a school district’s list of priorities. For simplicity, we can thus model an LEA’s
objective function as maximizing the aggregate achievement A within county c and year
t. As described above, each county has the authority to allocate yearly local revenues Rt
among a number of different purposes: providing teacher salary supplements, improving
buildings and facilities, paying for other certified and non-certified school personnel, and
paying for other operating expenses. In the basic model below, a county decides how
much to spend on non-instructional (NI) and instructional (I) categories in order to
maximize student achievement, assuming for the moment that local revenues to be used
toward education (R) are exogenously determined.
1 𝑚𝑎𝑥!!"!!,!!"!!
𝐴!" = 𝑓 𝑁!"!!, 𝐼!"!!,𝑿!"
𝑠. 𝑡. 𝑁!"!! + 𝐼!"!! ≤ 𝑅!"!! In this equation a vector of other county characteristics Xct also affects student
achievement, and counties are restricted by an annual budget constraint. I assume there is
a one-time lag period between school expenditures and student outcomes.
If I impose linearity and additive separability assumptions of achievement as a
function of the two types of school resources, this optimization problem becomes:
16
2 𝑚𝑎𝑥!!"!!,!!"!!
𝐴!" = 𝛽! + 𝛽!𝑁!"!! + 𝛽!𝐼!"!! + 𝜷!𝑿!" + 𝜇!"
𝑠. 𝑡. 𝑁!"!! + 𝐼!"!! ≤ 𝑅!"!! For this basic form of education production function, the solution would be:
𝐼!"!! =0 𝑖𝑓 𝛽! > 𝛽!
𝑅!"!! 𝑖𝑓 𝛽! ≥ 𝛽!; 𝑁!" =
0 𝑖𝑓 𝛽! ≥ 𝛽!𝑅!"!! 𝑖𝑓 𝛽! > 𝛽!
In words, counties would choose to allocate all revenues to whichever type of spending –
instructional or non-instructional – is found most effective for raising student
achievement. However, this situation is unrealistic for a number of reasons, including
federal and state requirements, interaction effects between the two types of spending, and
non-linearity in the production function in the form of diminishing returns.
The concept of fiscal substitution, popularized by research on state and local
behavioral responses to federal grant policy in the 1970s (Gramlich & Galper, 1973;
Johnson & Tomola,1977) and more recently in the context of school finance reform
(Baicker & Gordon, 2007), may be relevant toward understanding how county
governments react to state- or federal-imposed fiscal constraints. In North Carolina, the
strict statewide salary schedule and class size requirements leave counties few options for
supplementing classroom instruction through traditional teacher support. If non-
instructional support personnel are complementary to the benefits provided by teachers
(as modeled below), this may strengthen the case that local non-instructional resources
could benecit student learning.
To formalize this simple example in which we allow the impacts of instructional
and non-instructional resources to interact, the problem becomes:
3 𝑚𝑎𝑥!!",!!"
𝐴!" = 𝛽! + 𝛽!𝑁!" + 𝛽!𝐼!" + 𝛽!(𝐼!"×𝑁!")+ 𝜷!𝑿!" + 𝜇!"
𝑠. 𝑡. 𝑁!" + 𝐼!" ≤ 𝑅!"
17
This problem results in the solution below. Essentially, counties respond to both the
relative efficacy of each type of resource (𝛽! and 𝛽!), and the degree to which the
resource types interact (𝛽!):
𝑁!" , 𝐼!" =𝑅2 +
(𝛽! − 𝛽!)2𝛽!
,𝑅2 −
(𝛽! − 𝛽!)2𝛽!
Figure 2 provides a three-dimensional graphical representation of the associations
between log instructional per-pupil spending, log non-instructional per-pupil spending,
and student math and reading performance at the county-year level. Indeed, we observe
that the county-year observations with highest student test performance tend to have high
expenditures in both categories. On the other end of the spectrum, there also seem to be
counties performing well on standardized tests but with low levels of overall per-pupil
spending. This could perhaps be a “large district effect” wherein larger districts are able
to provide certain services at a lower per-pupil cost due to economies of scale, although
there are numerous alternative explanations.
Without a causal estimation approach, it is impossible to tease out the direct effect
of county education spending priorities on student outcomes. In contrast to school
boards, county commissioners have an array of interrelated objectives, including
promoting the safety, health, and economic prosperity of the entire county population.
Therefore their goals are not as simple as in the decision-making model above, and they
are likely to disagree on the relative effectiveness of different investments. The relative
benefits and costs of non-instructional and instructional resources to student performance,
as estimated causally in this study, could therefore be informative for policy-makers at all
levels of government as they consider how to optimize returns to expenditures.
18
My identification strategy, detailed in the next section, arises directly from the
process by which county governments raise revenues and allocate educational and non-
educational expenditures across different functions each year. In particular, due to the
fairly high level of discretion of county commissioners in the budgetary process, I can
take advantage of reliable sources of exogenous variation in instructional and non-
instructional resources at the local governmental level.
4. Empirical approach
Data
This study introduces a new dataset consisting of student-level administrative data
in North Carolina merged with detailed school expenditures data by purpose, type of
expenditure, and revenue source. The dataset includes all 100 counties tracked for 18
years between 1996 and 2013, matched to over 13 million student observations from
grades K-12 during this period. This comprehensive time coverage and large student
sample size allow for very precise estimates of how local education expenditures affect
students across a number of domains. Table 1 depicts descriptive statistics, including
demographic characteristics, parental education, and educational indicators, of a cross-
section of the main student dataset.
Student achievement outcomes for this study include End of Grade test scores in
reading and math for students in grades three through eight, and End of Course test scores
for students from grades nine through twelve in English I; English II; U.S. History;
Economics, Law, and Politics; Civics; Biology; Physical Science; Chemistry; Physics;
Algebra II; and Geometry. All test scores are normalized by grade level, year, and subject
19
to have mean zero and standard deviation one. Therefore test score results in the
following sections are presented as effect sizes in terms of test score standard deviations.
For students in grades nine through twelve, I average all normalized science and math
scores for a given student in a given year to construct an annual “sciences” score. I do
the same with English and social studies course scores to construct a “humanities”
measure for each student-year observation.7 Combining across years and grades, the
dataset contains approximately 7.3 million observations for End of Grade scores and 2.5
million observations for End of Course scores.
Recent empirical research has demonstrated the long-term importance of non-
academic personal attributes for educational attainment, labor market outcomes, and
social outcomes (Heckman, Stixrud, & Urzua, 2006). Importantly, a series of recent
studies on teacher quality have shown that administrative behavioral measures can
reliably capture student non-cognitive traits and behaviors (Gershenson, Forthcoming;
Jackson, 2012; Ladd & Sorensen, Forthcoming). Therefore, to the extent possible using
administrative data, this study seeks to measure outcomes indicative not only of
curricular learning but also of non-academic behaviors and traits. I estimate the effects of
expenditures on the following non-test outcomes: attendance, disciplinary infractions,
and self-reported extracurricular participation. Student attendance is measured by a
continuous measure of absences during the school year. A disciplinary infraction measure
equals one if the student received in-school suspension, out-of-school suspension,
expulsion, or placement in an alternative school or alternative learning program that year.
7 These scores are computed as row means such that the “sciences” score of students with only one sciences/math course score in a given year will be that score and such that the “sciences” score of students with several science/math course scores in that year will be an average of all normalized scores.
20
It equals zero otherwise. And an extracurricular participation measure likewise equals
one if a student in ninth through twelfth grade reports participating in an academic,
athletic, artistic, community service, or vocational club in the current academic year. This
measure equals zero if students report that they did not participate in any of these types of
activities that year.
I have matched these student-level outcomes to education expenditures
information from the North Carolina Department of Instruction, also dating back to the
1995-1996 school year. I observe county annual current expenditures, categorized by
purpose code and object code. Accountants use the four-digit purpose codes to designate
how much money schools and districts spend on particular functions.8 The object codes
on the other hand signify the type of spending: salaries, employee benefits, purchased
services, supplies and materials, or instructional equipment. For the purpose of this
analysis, I include only salary expenditures because they reflect the direct effect of
increasing personnel in school and because they are measured more consistently over
time than other categories such as benefits. I adjust all expenditures data for inflation to
2013 equivalent dollars.
Also at the county-year level, this dataset contains a comprehensive set of time-
varying covariates: population estimates and poverty levels from the U.S. Census Bureau;
the local unemployment rate from the Bureau of Labor Statistics; political party
affiliation rates from the State Board of Elections; and yearly total county expenditures
and property assessment values from the County Budget and Tax Survey. These
8 Instructional spending includes salary expenditures in the purpose code 5100 (Regular instructional programs); Non-instructional spending includes salary expenditures in the following purpose codes: 5820 (Attendance – social work services); 5830 (Guidance services); and 5840 (Health services). Exact descriptions of these categories are provided in Appendix Table A1.
21
economic, political, and population change measures act mostly as control variables in
the estimation framework described below.
Identification strategy
The purpose of this study is to estimate the effects of local non-instructional and
instructional expenditures on student outcomes in grades three through twelve, including
test scores across many subjects, absences, disciplinary infractions, and extracurricular
participation. A standard OLS regression of outcomes on the mix of spending would
likely lead to biased estimates since the county’s level of per-pupil expenditures on
various purposes is endogenous (Haegeland, Raaum, & Salvanes, 2012; Hanushek
1997).9 Three main types of omitted variables likely introduce bias to a regression
model: 1) observable and unobservable characteristics of counties; 2) time trends and
contemporaneous events; and 3) time-varying local economic, political, or population
factors.
To account for the first two sources of endogeneity, I include county and year
fixed effects in all models. This implies that my estimates reflect the effects of within-
county changes in local education spending across time. However, county time-varying
characteristics could still drive both local spending patterns and student achievement
even with county and year fixed effects. For example, counties could respond to local
financial conditions or to problems and shortages as they arise in their schools by
9 Appendix Table A2 compares IV estimates to traditional OLS estimates for returns to instructional and non-instructional spending. OLS estimates of the effects of education spending appear to be downward biased, consistent with the conclusions of some prior research (e.g. Haegeland, Raaum, & Salvanes, 2012). One plausible explanation is that county commissioners allocate additional instructional or non-instructional funds in response to perceived need of students, which would lead to problems of simultaneity.
22
increasing education spending in certain categories, but the underlying conditions could
also affect student learning and behaviors.10 Therefore, in addition to the county and year
fixed effects, I employ an instrumental variables approach to account for the nonrandom
allocation of resources within counties across years. The model is further complicated by
the fact that it includes three endogenous variables on the right hand side of the equation:
non-instructional spending, instructional spending, and total county expenditures.
Therefore I need three instrumental variables to provide enough exogenous variation in
local expenditures. For this task I have identified: first, a measure of how recently
property value reassessment took place in the county; second, a measure of the political
party majority of the county board of commissioners; and third, interaction terms of these
two measures. I describe and justify these three exogenous sources of variation in more
detail below.
Revaluation instrument
North Carolina General Statutes require counties to reassess real property values
every eight years, or more frequently if the county so chooses. Each county operates on a
different revaluation schedule, and in revaluation year counties often experience a
noticeable jump in increased property tax revenue (or a dip if average property values
have decreased since the previous revaluation year). Walden (2003) and Ladd (1991)
observe this phenomenon in North Carolina counties, and Bloom and Ladd (1982) found
10 Findings from a simple regression of education expenditures on county economic conditions suggest that this is likely the case. County-level unemployment claims (lagged by two years) significantly affect education expenditures even when county and year fixed effects are included. The literature on this topic finds that local job loss has detrimental effects on population-level student achievement (Ananat, Gassman-Pines, Francis, & Gibson-Davis, 2011). Therefore omitting key local economic indicators could introduce significant bias into OLS estimates of the effects of county education expenditures on student outcomes.
23
that Massachusetts counties also tend to increase their property tax revenues during
revaluation years.
As seen in Figure 3a, from the recent North Carolina data from 1996 to 2013,
county revenues from property taxes increase by over 40 percent in the year after
revaluation on average. Furthermore, this sudden increase in revenues translates to the
amount that counties ultimately spend on education and other purposes, although the
effect is typically lagged one year (See Figure 3b). I construct an instrumental variable
that equals the number of years since revaluation, and a corresponding measure that
includes the number of years since revaluation squared. An illustration of how these
measures work in practice is provided in Appendix Table A2 for three fictional counties
over a ten year period.
The revaluation instrument as described above significantly predicts per-pupil
total county expenditures and spending on non-instructional and instructional functions
even when controlling for county and time fixed effects and a set of county time-varying
control variables related to population changes and economic conditions. Table 2
presents first stage estimates from this instrument, the political instrument described
below, and interaction terms of each instrumental variable (F1=6,060; F2=9,946;
F3=14,530).11 It is possible that counties begin increasing education expenditures before
revaluation years in anticipation of property values increasing, or that they choose to set
the tax rate in the revaluation year at a “revenue-neutral rate” such that revenue flows
11 These F-statistics come from first stage estimates in the sample of students in grades 3-8, including student covariates in addition to the county covariates. First stage estimation was replicated in other student samples and at the aggregated county-year level with only 1,800 observations. All instrumental variables remain strong predictors of county spending, even in these alternative samples.
24
remain constant. But in either of these instances, one would expect the effect of property
revaluation on local expenditures to be slightly downward biased.
Political instrument
Political party majority affiliation of the county board of commissioners in the
two years prior to measured school expenditures serves as the second instrumental
variable for this analysis. The county board of commissioners is primarily responsible
for determining annual budgetary allocations and setting local tax rates. These boards
contain either five or seven members, each with four-year terms, and elections are held
every two years. For this reason, the political party majority of a board often switches
back and forth within counties across years.12 In Figure 4, a local polynomial smooth
plot of per-pupil county expenditures (with the 95 percent confidence interval plotted in
gray) shows that for several years preceding political party turnover events, expenditures
are quite stable. However, upon party majority change – in this case from a Republican
to Democratic majority – county expenditures significantly increase. As can be seen in
Figure 5, which graphs the percent of registered voters that are Democrats, political party
turnover events do not appear to be caused by any underlying major political shifts.
This political instrumental variable significantly predicts both the total level of
per-pupil educational spending as well as the proportion allocated to instructional and
non-instructional purposes (See Table 2). As shown in Table 2, Republican-majority
county commissioners tend to spend less in general and less on non-instructional
investments, but more on instructional salary supplements. I assert that the measure of
Republican party majority of county commissioners is unlikely, given the inclusion of 12Sixty-three percent of counties switch political party majority during the observed time period.
25
both county and year fixed effects, to relate to student and youth outcomes through any
mechanisms other than public expenditure. I also control for the annual proportion of
voters registered as Republicans so that, conditional on any underlying political shifts, the
county commissioner political instrument provides truly exogenous variation in local
spending.
Interaction term of political and revaluation instruments
As a third source of exogenous variation, I include two interaction terms of the
two revaluation instrumental variables each multiplied by the political instrumental
variable indicator. As presented numerically in Table 2 and depicted graphically in
Figures 6a through 6c, predicted expenditures have much steeper gradients by years since
property revaluation for Democratic-majority boards of county commissioners than for
Republican-majority boards of county commissioners. This finding implies that
Democratic boards are more likely to take advantage of the timing of property value
reassessment to increase total county revenues. As with the other two identified sources
of variation in local education expenditures, these interaction terms appear to meet all
necessary criteria for instrumental variables (having no correlation with the error term,
and showing strong correlation with the endogenous predictors of interest).
Estimation model
To estimate the effects of non-instructional and instructional salaries on student
outcomes, I use full information generalized two stage least squares estimates of the
system of simultaneous equations (Balestra & Varadharajan-Krishnakumar, 1987). The
26
first stage equations, which predict total county expenditures (minus the two other
categories) Tc,t; non-instructional salaries Nc,t; and instructional salaries Ic,t; for county c
and year t are as follows:
(1a)𝑇!,! = 𝛽! + 𝛽!𝑃𝑜𝑙𝑖𝑡𝐼𝑉!,!!! + 𝜷𝟏𝑹𝒆𝒗𝒂𝒍𝑰𝑽𝒄,𝒕!𝟏 + 𝜷𝟐(𝑃𝑜𝑙𝑖𝑡𝐼𝑉!,!!!×𝑹𝒆𝒗𝒂𝒍𝑰𝑽𝒄,𝒕!𝟏)+𝜷𝟑𝑿𝒄,𝒕 + 𝜃! + 𝛿! + 𝜇!,!
(1b)𝑁!,! =
𝛽! + 𝛽!𝑃𝑜𝑙𝑖𝑡𝐼𝑉!,!!! + 𝜷𝟏𝑹𝒆𝒗𝒂𝒍𝑰𝑽𝒄,𝒕!𝟏 + 𝜷𝟐(𝑃𝑜𝑙𝑖𝑡𝐼𝑉!,!!!×𝑹𝒆𝒗𝒂𝒍𝑰𝑽𝒄,𝒕!𝟏)+𝛽!𝑋!,! + 𝜃! + 𝛿! + 𝜇!,!
(1c) 𝐼!,! = 𝛽! + 𝛽!𝑃𝑜𝑙𝑖𝑡𝐼𝑉!,!!! + 𝜷𝟏𝑹𝒆𝒗𝒂𝒍𝑰𝑽𝒄,𝒕!𝟏 + 𝜷𝟐(𝑃𝑜𝑙𝑖𝑡𝐼𝑉!,!!!×𝑹𝒆𝒗𝒂𝒍𝑰𝑽𝒄,𝒕!𝟏)+
𝛽!𝑋!,! + 𝜃! + 𝛿! + 𝜇!,! Expenditures for these three separate purposes are thus regressed on the political
instrument (𝑃𝑜𝑙𝑖𝑡𝐼𝑉!,!!!), the vector of revaluation instruments (𝑹𝒆𝒗𝒂𝒍𝑰𝑽𝒄,𝒕!𝟏), and
interaction terms between those measures (𝑃𝑜𝑙𝑖𝑡𝐼𝑉!,!!!×𝑹𝒆𝒗𝒂𝒍𝑰𝑽𝒄,𝒕!𝟏). These first stage
models also include county fixed effects (𝜃!), year fixed effects (𝛿!), and county time-
varying indicators (𝑿𝒄,𝒕) including population estimates, student-age poverty rates,
unemployment rates, assessed property values, registered voter affiliation rates, and
average student characteristics (gender, race/ethnicity, and prior test performance).
From the first stage system of equations, with three endogenous spending
variables and three exogenous instruments, I can therefore predict values for local
expenditures derived from variation in years since property revaluation, political majority
of county government, and interactions between the two. In the second stage, I estimate
the effect of lagged predicted spending in several categories on a student outcome 𝑌!"#:
(2) 𝑌!,!,! = 𝛽! + 𝛽!𝑁!,!!! + 𝛽!𝐼!,!!! + 𝛽!𝑇!,!!! + 𝜷𝟒𝑿𝒊,𝒄,𝒕 + 𝜃! + 𝛿! + 𝜀!,!,!
In this equation, I regress each of student i’s outcomes in year t county c on predicted
non-instructional, instructional, and other county-level spending in the period prior
27
(𝑁!,!!!; 𝐼!,!!!; and 𝑇!,!!!); a vector of student-level covariates including prior test scores,
gender indicators, and race/ethnicity indicators (𝑿𝒊,𝒄,𝒕); and county-level covariates as in
the first-stage equations. Once again, I include county and year fixed effects such that I
exploit only within-county variation over time in education resource allocation.
The timing of how quickly expenditures may affect student learning and
behaviors, as well as the timing of when property revaluation and political shifts
influence local expenditures, are both critical to this model. This analysis assumes that
student outcomes in a certain year are a function of average expenditures in the two years
prior to measurement. Similarly, this study assumes that a one-year lag exists between
local expenditures and their prior predictors (county government political majority
changes and property value reassessment). I perform a number of robustness tests to
ensure that findings in both the first and second stages are not sensitive to the choice of
time lags.13
5. Results
Answering the question of how non-instructional investments affect student
outcomes could inform how schools and local, state, and federal governments in the
future choose to allocate resources effectively. The two-stage model described in the
previous section allows me to estimate the causal impact of increased spending on
teachers and teaching assistants or school social workers, guidance counselors,
psychologists, and health services. I estimate these impacts across a number of domains
13 I experiment with other lag periods for expenditures to outcomes (0-1 years; 1-3 years; 2-3 years; 3-5 years), and determine that the direction and magnitude of findings are similar when the time lag is anywhere between 1 and 3 years, and nonsignificant with smaller or larger lags.
28
of student performance: math and reading scores in grades 4-8; sciences and humanities
performance in grades 9-12; absences per year; serious disciplinary infractions; and
extracurricular participation in grades 9-12.
Table 2 presents results from the first stage equations. In terms of the property
revaluation instruments, we can note that expenditures in all spending categories increase
shortly after a property value reassessment year, and then decline steadily in the years
following that. Figures 6a, 6b, and 6c show these quadratic trends from the first stage
estimates graphically (separately for Republican-majority and Democratic-majority
boards). One can observe that Republican-majority boards of county commissioners
spend less per-pupil than do Democrat-majority boards of county commissioners. They
also allocate less money per-pupil on non-instructional expenses, although they allocate
more money toward instructional salaries. Interestingly, Republican-majority boards
appear to respond less dramatically to the property revaluation schedule. One could
conclude that they have lower propensity to use the revaluation timing as an opportunity
to increase property tax revenues and public expenditures.
Student achievement outcomes
Moving to second-stage estimation, first I look at the impact of increased non-
instructional and instructional expenditures on student test scores. As can be seen in
Table 3, a $100 increase in per-pupil expenditures on non-instructional services (social
workers, guidance and psychological services, and health services) leads to a 0.067
standard deviation increase in math scores and a 0.014 standard deviation increase in
reading scores in grades 4 through 8. Given the relatively small investment of money,
29
this magnitude of the test score increase is meaningful. The math effect, for example, is
equivalent to one-fifth of the total average difference between sixth grade and seventh
grade math performance. The same dollar increase in instructional salary spending
induces somewhat smaller, but statistically significant, improvements in test performance
(0.016 SD for math and 0.029 for reading).
At the high school level, a $100 per-pupil increase in non-instructional spending
translates to increases in average science and math course performance of 0.099 standard
deviations but no increases in average humanities and social sciences course performance
(See Table 3). The same $100 per-pupil increase in local instructional salary
supplements augments science and math performance in high school by 0.078 standard
deviations. Instructional salary spending also does not affect average humanities and
social sciences performance. The results at the high school level parallel those found at
the elementary school and middle school levels, suggesting that these investments do not
have widely varying impacts by grade level.
As discussed in the methods section, the omission of other county expenditures
from the model (whether educational or non-educational) could bias upwards estimated
effects of increasing instructional or non-instructional salaries. Therefore, I include in the
model total other county expenditures as an endogenous regressor simultaneously with
non-instructional and instructional salaries. As can be seen in Tables 3 and 4, total
county expenditures on other categories appear to have little to no effect on student test-
based or behavioral outcomes. Surprisingly, increasing total county expenditures appears
to have a small but statistically significant negative effect on student reading scores
(Table 3) and positive effect on student absenteeism (Table 4). Although the point
30
estimate magnitudes are small enough to not be too concerning, I cannot conclude
definitively what mechanisms may be generating these results.
The logic behind instructional spending’s impact on student achievement is clear.
Instructional spending can pay for more experienced teachers and attract higher quality
teachers, and it can also subsidize teacher assistants to help in the classroom. These
investments effectively raise student performance, particularly in quantitative subjects.
Why non-instructional spending would have such significant effects on student test
scores is less apparent. The multiple instrumental variables technique identifies effects of
the two endogenous expenditures measures separately, and so collinearity between non-
instructional and instructional spending is unlikely to explain the effect. Providing
students with access to guidance, psychological, social work, and health services may
allow students to focus and participate more fully in classroom learning. To test potential
behavioral mechanisms of the effect, I examine the relation between non-instructional
resources and student absences, disciplinary infractions, and extracurricular participation.
To double-check the robustness of the identification strategy, I test for reverse
causality using a strategy similar to that used by Rothstein (2009) to identify bias in
teacher value-added models. Specifically, I estimate the impact of increasing
instructional or non-instructional salary spending in year t+1 on student outcomes in year
t, applying the same instrumental variables approach and control covariates as before.
Of the eight outcomes measured in grades four through eight, the main per-pupil
spending variables affect only one with marginal significance (See Appendix Table A4).
Other than this one negative impact on reading achievement (significant at the 10%
31
confidence level), the test provides no evidence that other underlying factors may be
driving this study’s main findings.
Student behavioral outcomes
As with student test scores, I estimate the effects of investments in social workers,
guidance and psychological services, and health services on student behavioral outcomes.
Jackson (2014) demonstrates that behavioral outcomes measured in student
administrative data robustly predict adult economic and social outcomes. The first
outcome of interest is student absences. Reported in Table 4, an additional $100 per
pupil on non-instructional services reduces average student absences by 0.57 days per
year (down from the average number of absences for the typical student of 8 days).
Instructional spending of the same quantity also decreases student absenteeism, but by a
smaller amount (0.03 absences). These results are highly policy-relevant because
frequent absence, even as early as the sixth grade, reliably predicts the likelihood of a
student eventually dropping out of school (Allensworth & Easton, 2007; Balfanz, Herzog,
& Mac Iver, 2007).
For disciplinary infractions, a $100 per-pupil increase in non-instructional
spending leads to a 0.98 percentage point decline in the proportion of students who
receive at least one suspension, expulsion, or placement in an alternative learning
program or school during the school year. This decrease is from a baseline rate of 10% of
students incurring such serious disciplinary consequences each year. And finally, for
extracurricular participation in high school, a $100 increase in per-pupil non-instructional
spending produces a 13.9% increase in self-reported participation in academic clubs,
32
sports, arts, community service, or vocational clubs, up from a baseline level of 65%.
Instructional spending has no statistically significant impact on student disciplinary
infractions or on self-reported extracurricular participation. (The non-significant
coefficients on instructional spending are -0.012 for disciplinary infractions and 0.007 for
extracurricular participation.) Table 4 presents main findings for all behavioral
outcomes.
Up to this point I have framed the student behavioral indicators as potential
mediators of the effect of non-instructional resources on student achievement, but I have
not tested this mediation directly. To do so, I can perform structural equation model
(SEM) estimation and through this process tease out direct and indirect effects of
increasing instructional and non-instructional expenditures on student test scores. In the
case of continuous outcome variables and no latent (unobserved) variables, SEM
estimates provide identical results to a three-staged reduced form treatment effect
decomposition used by Heckman, Pinto, and Savelyev in an evaluation of the Perry
Preschool program and other similar research (Heckman, Pinto, & Savelyev, 2013;
Sorensen, Dodge, and the Conduct Problems Prevention Research Group, Forthcoming).
Figures 7a and 7b illustrate the path model for which maximum likelihood
estimation is conducted, correcting for missing data and controlling for student
covariates. This method, unlike the rest of the analysis performed for this paper, relies on
variable covariance rather than on quasi-experimental effects and therefore should be
interpreted with caution. The figures present estimated direct path coefficients for each
one-directional path and indicate that, as expected, increased expenditures are associated
with reduced student absences and disciplinary infractions, which in turn then predict
33
student test scores. Combining direct and indirect effects estimated by the model, non-
academic behaviors explain 30.7 percent of the total effect of instructional spending and
88.6 percent of the total effect of non-instructional spending on math test scores. For
reading achievement, non-academic behaviors account for 19.2 percent of the overall
instructional spending effect and 36.0 percent of the non-instructional spending effect.
Again, this model does not pass the strict exogeneity assumption, but the findings
interestingly suggest that absences and disciplinary infractions play a large role in the
relation between non-instructional salaries and student test scores (and a relatively larger
mediation role than for instructional expenditures).
Heterogeneity by school poverty
In general, these non-instructional services may hold value due to the extent that
they assist students with physical, mental, economic, and personal obstacles to learning.
With this hypothesis in mind, I test whether non-instructional spending has greater
impact in areas with high concentrations of poverty versus areas with low concentrations
of poverty. I define high-poverty counties as those with a student-age poverty rate
greater than 25 percent for most years in the observed sample. Likewise, low-poverty
counties are those with student-age poverty rates below 25 percent for a majority of
years. Table 5 provides estimates of the effects of non-instructional services on student
achievement in grades 4 through 8 for the sample, split by county poverty level.
A $100 increase in 2013 dollars of per-pupil non-instructional spending raises
math achievement by 0.012 standard deviations and reading achievement by 0.004
standard deviations for low-poverty counties (n=5.5 million). This impact is much higher
34
in the high-poverty counties, with increased non-instructional spending associated with
0.239 standard deviations growth in math and 0.077 standard deviations growth in
reading (n=1.9 million). A similar pattern arises for instructional salaries. In low-
poverty counties, a $100 per-pupil investment in instructional salary supplements
increases math scores by 0.004 standard deviations and reading scores by 0.006 standard
deviations; in high-poverty counties these values are 0.065 standard deviations and 0.098
standard deviations. (The math effect for instructional expenditures in high-poverty
counties is not significant.)
Returns to non-instructional spending are nearly 20 times greater for math
performance in high-poverty areas than low-poverty areas, with similar patterns emerging
from reading score results. These estimates suggest that investments in school support
personnel such as social workers, health services, and guidance counselors, in schools
with concentrated low-income student populations could go a long way toward closing
socioeconomic achievement gaps.
Nonlinearity in expenditure effects
Thus far all estimates have represented the average linear effect of supplementing
either instructional or non-instructional resources by 100 additional dollars per pupil.
However, the baseline levels of spending toward instructional and conversely non-
instructional purposes are quite disparate. In 2013, the average North Carolina county
spent $2,491 per pupil on instructional salaries but only $228 per pupil on non-
instructional salaries. Therefore an exogenous $100 boost in expenditures would alter the
non-instructional budget much more dramatically than it would the instructional budget.
35
Furthermore, if there are diminishing marginal returns to education spending, then the
relatively large effects for increasing non-instructional spending may not be so surprising.
To address some of these issues, I explore the possibility of non-linearity in
returns to local education expenditures. To do so, I first estimate the first stage models as
presented in equations 1a, 1b, and 1c, and predict local non-instructional, instructional,
and total county expenditures. For convenience, I focus on just one of the (most salient)
outcomes from the main results: math test score achievement in grades 4 through 8. For
both the math performance outcome and the predicted local spending variables, I demean
each measure by county and year. At this point, I can perform kernel-weighted local
polynomial regression to learn more regarding the non-parametric relationship between
predicted non-instructional and instructional spending and student achievement.
In Figures 8a, 8b, and 8c, I present graphically how predicted local instructional,
non-instructional, and total county spending from the first stage estimates relates to math
performance, as measured in standard deviations from the mean. The marginal effect of
spending more per pupil on local instructional salaries (Figure 8a) appears to be fairly
linear between $0 and $1,500. These estimates are more precise for low levels of
instructional salary spending because a higher number of county-year observations exist
in that range. For non-instructional spending (Figure 8b), an interesting picture emerges.
Any spending within the range of $0 to $50 per pupil appears to have little measurable
effect on student learning; however, moving from spending $50 to spending $100 per
pupil increases math performance by 0.2 standard deviations – a large, rapid increase.
Above $100, the diminishing marginal returns set in and more spending does not lead to
significant math score raises.
36
These non-parametric estimates of the returns to non-instructional resources are
perhaps more informative than the linear estimates. We can observe that the large impact
of providing additional social and health services in schools estimated in this study may
be an artifact of very low baseline spending on these student resources. A small
investment at the county level in non-instructional expenditures (of roughly $100 per-
pupil) could yield large benefits for students, particularly those in high-poverty counties,
but spending above that threshold may not produce any consequential gains. Figure 8c
graphs predicted total county expenditures per pupil (including education expenditures)
against math performance. Although there is a small positive linear effect in the lower
ranges of county expenditures, larger amounts of county expenditures only generate a
noisy net zero effect.
6. Discussion
The primary spending categories examined in this study are of public interest
even without any connection to student learning. Social workers, guidance counselors,
and psychologists provide needed support for children who may be experiencing
educational, mental health, family, or social problems. Health services such as school
nurses, health centers, and dental care, can likewise improve the physical health and
health behaviors of students. However, this study is the first to demonstrate at the
aggregate level that spending on these non-traditional educational categories may
facilitate student learning and improve other behavioral indicators (such as attendance
and disciplinary offenses) predictive of long-term educational success.
37
Corresponding inference about the impact of instructional spending in schools is
limited by the constraints of this study design, due to the fact that I can only look at
effects of instructional salary supplements and of hiring teaching assistants rather than
effects of hiring more teachers per pupil, for example. Local instructional spending,
which again is composed primarily of teacher salary supplements and hiring of teaching
assistants, still appears to modestly boost student learning even in the relative short-term.
The inconsistencies of results in prior school finance research could in part reflect
large heterogeneities in returns to school resources by the type of school resource. For
example, Greenwald, Hedges, and Laine (1996) conclude that the median effect of a $100
increase in per-pupil education expenditures, across an array of former studies, is a 0.002
standard deviation increase in student achievement. The current study, which uncovers
effect sizes larger by an order of magnitude, finds that the marginal effect of increasing
spending on non-instructional or instructional salaries in schools greatly exceeds prior
estimates of returns to school spending, which typically aggregated across a large set of
categories (including for example administration at all levels, retirement and benefits,
capital expenses, and materials). Further research that delves more into the relative
effectiveness of different types of educational spending could provide valuable insight.
Given the serious and growing prevalence of mental health, health, and social-
emotional concerns that K-12 public school students face, perhaps the positive benefits of
school support services are not so surprising. Across the country, states and school
districts are experimenting with “community school” models in which students are
provided comprehensive social, behavioral, and health services, alongside their classroom
activities. A new model in which schools are not only places of instruction but also
38
providers and protectors of social well-being and health of students may in fact make
progress toward closing achievement gaps between low-socioeconomic-status and high-
socioeconomic-status children and youth. Further research on this topic could better
tease out the mechanisms through which non-instructional spending produces change in
students, investigate long-term effects on educational attainment and social outcomes,
and disaggregate the total non-instructional category to assess its components (social
workers and attendance, health services, and guidance counselors and psychologists)
individually.
39
References Allensworth, E. M. & Easton, J. Q. (2007). What matters for staying on-track and
graduating in Chicago Public Schools. Chicago, IL: Consortium on Chicago
School Research at the University of Chicago.
Allen-Meares, P., Montgomery, K. L., & Kim, J. S. (2013). School-based social work
interventions: a cross-national systematic review. Social Work, 58 (3), 253-262.
Ananat, E. O., Gassman-Pines, A., Francis, D. A., Gibson-Davis, C. M. (2011). Children
left behind: Effects of statewide job loss on student achievement. NBER Working
Paper 17104.
Baicker, K. & Gordon, N. (2006). The effect of state education finance reform on total
local resources. Journal of Public Economics, 90 (8-9), 1519-1535.
Balestra, P. & Varadharajan-Krishnakumar, J. (1987). Full information estimations of a
system of simultaneous equations with error component structure. Econometric
Theory, 3 (2), 223-246.
Balfanz, R., Herzog, L. & Mac Iver, D. J. (2007). Preventing student disengagement and
keeping students on the graduation path in urban middle-grades schools: Early
identification and effective interventions. Educational Psychologist, 42 (4), 223-
235.
Bloom, H. S. & Ladd, H. F. (1982). Property tax revaluation and tax levy growth.
Journal of Urban Economics, 11 (1), 73-84.
Bureau of Labor Statistics (BLS). (2015). Consumer Price Index – All Urban Consumers.
Series ID: CUUR0000SA0. U.S. City Average All Items 1996-2014.
40
Carrell, S. E. & Hoekstra, M. (2011). Are school counselors a cost-effective education
input? Economics Letters, 125 (1), 66-69.
Centers for Disease Control and Prevention (CDC). (2013). Mental health surveillance
among children – United States, 2005-2001. Morbidity and Mortality Weekly
Report Supplements, 62 (2), 1-35
Currie, J., Stabile, M., Manivong, P., & Roos, L. L. (2010). Child health and young adult
outcomes. Journal of Human Resources, 45 (3), 517-548.
McMurrer, J. (2007). NCLB Year 5: Choices, changes and challenges: Curriculum and
instruction in the NCLB Era. Washington, DC: Center on Education Policy.
Dee, T. S. (2005). Expense preference and student achievement in school districts.
Eastern Economic Journal, 31 (1), 23-44.
Deluca, T. A. (2015). Do countywide LEAs allocate expenditures differently from
community-centric LEAs? Evidence from National Center for Education Statistics
Common Core Data. Journal of Education Finance, 40 (3), 222-252.
Feldman Ferb, A. F. & Matjasko, J. L. (2012). Recent advances in research on school-
based extracurricular activities and adolescent development. Developmental
Review, 32 (1), 1-48.
Gershenson, S. (2016). Linking teacher quality, student attendance, and student
achievement. Education Finance and Policy, 11 (2), 1-22.
Gifford, E. J, Wells, R., Bai, Y., Troop, T. O., Miller, S., & Babinski, L. M. (2010).
Pairing nurses and social workers in schools: North Carolina’s school-based child
and family support teams. Journal of School Health, 80 (2), 104-107.
41
Gramlich, E. M. & Galper, H. (1973). State and local fiscal behavior and federal grant
policy. Brookings Papers on Economic Activity, 1, 15-58.
Greenwald, R., Hedges, L. V., & Laine, R. D. (1996). The effect of school resources on
student achievement. Review of Educational Research, 66 (3), 361-396.
Haegeland, T., Raaum, O., & Salvanes, K. (2012). Pennies from heaven? Using
exogenous tax variation to identify effects of school resources on pupil
achievement. Economics of Education Review, 31 (5), 601-614.
Hanushek, E. A. (1997). Assessing the effects of school resources on student
performance: An update. Educational Evaluation and Policy Analysis, 19 (2),
141-164.
Heckman, J. J., Pinto, R., & Savelyev, P. (2013). Understanding the mechanisms through
which an influential early childhood program boosted adult outcomes. American
Economic Review, 103 (6), 2052-2086.
Heckman, J. J., Stixrud, J., & Urzua, S. (2006). The effects of cognitive and non-
cognitive abilities on labor market outcomes and social behavior. Journal of
Labor Economics, 24 (3), 411-482.
Hille, A. & Schuppe, J. (2015). How learning a musical instrument affects the
development of skills. Economics of Education Review, 44, 55-82.
Jackson, C. K. (2014). Non-cognitive ability, test scores, and teacher quality: Evidence
from 9th grade teachers in North Carolina. NBER Working Paper No. 18624.
42
Jackson, C. K., Johnson, R., & Persico, C. (2015). The effects of school spending on
educational and economic outcomes: Evidence from school finance reforms.
NBER Working Paper No. 20118.
Johnson, G. E. & Tomola, J. D. (1977). The fiscal substitution effect of alternative
approaches to public service employment policy. Journal of Human Resources,
12 (1), 3-26.
Ladd, H. F. (1991). Property tax revaluation and tax levy growth revisited. Journal of
Urban Economics, 30 (1), 83-99.
Ladd, H. F. & Sorensen, L. C. (Forthcoming). Returns to teacher experience: Student
achievement and motivation in middle school. Education Finance and Policy.
Lovenheim, M. F., Reback, R., & Wedenoja, L. (2014). How does access to health care
affect health and education? Evidence from school-based health center openings.
Working Paper.
Maughan, E. & Troup, K. D. (2011). The integration of counseling and nursing services
into schools: A comparative review. The Journal of School Nursing, 27 (4), 293-
303.
Mensah, Y. M., Schoderbek, M. P., & Sahay, S. P. (2013). The effect of administrative
pay and local property taxes on student achievement scores: Evidence from New
Jersey public schools. Economics of Education Review, 34, 1-16.
National Center for Education Statistics (2013). Common Core of Data.
North Carolina Center for County Research. (2015). Basics of county funding of public
schools. Raleigh, NC: North Carolina Association of County Commissioners.
43
North Carolina Department of Public Instruction (NC DPI). (2007). Uniform Chart of
Accounts. Prepared by Division of School Business.
North Carolina General Statutes. §115C-408 Elementary and Secondary Education:
Funds under control of the State Board of Education (2014).
Pfeifer, C., & Cornelißen, T. (2010). The impact of participation in sports on educational
attainment—New evidence from Germany. Economics of Education
Review, 29(1), 94-103.
Reardon, S. F. (2011). The widening academic achievement gap between the rich and the
poor: New evidence and possible explanations. In R. Murnane & G. J. Duncan
(Eds.), Whither Opportunity (91-116). New York, NY: Russell Sage Foundation.
Reback, R. (2010a). Noninstructional spending improves noncognitive outcomes:
Discontinuity evidence from a unique elementary school counselor financing
system. Education Finance and Policy 5 (2), 105-137.
Reback, R. (2010b). School’s mental health services and young children’s emotions,
behavior, and learning. Journal of Policy Analysis and Management, 29 (4), 698-
725.
Rothstein, J. (2009). Student sorting and bias in value-added estimation: Selection on
observables and unobservables. Education Finance and Policy, 4 (2), 537-571.
Sorensen, L. C., Dodge, K. A., & CPPRG. (Forthcoming). How does the Fast Track
intervention prevent adverse outcomes in adulthood? Child Development.
44
U.S. Department of Education, National Center for Education Statistics. (2013). National
Public Education Financial Survey, 1990-91 through 2010-11. Common Core of
Data (CCD).
U.S. Department of Health and Human Services. (1999). Mental health: A report of the
Surgeon General. Rockville, MD: U.S. Department of Health and Human
Services, Substance Abuse and Mental Health Services Administration, Center for
Mental Health Services, National Institutes of Health, National Institute of Mental
Health.
U.S. Public Health Service. (2000). Report of the Surgeon General’s Conference on
Children’s Mental Health: A National Action Agenda. Washington, DC:
Department of Health and Human Services.
Verstegen, D. A. & King, R. A. (1998). The relationship between school spending and
student achievement: A review and analysis of 35 years of production function
research. Journal of Education Finance, 4 (2), 243-262.
Walden, M. L. (2003). Improving revenue flows from the property tax. Popular
Government, 69 (1), 13-17.
Walden, M. L. & Sogutlu, Z. (2001). Determinants of intrastate variation in teacher
salaries. Economics of Education Review, 20 (1), 63-70.
45
Figures and Tables Figure 1a. North Carolina per-pupil instructional salary spending: 1996-2013
Figure 1b. North Carolina per-pupil non-instructional salary spending: 1996-2013
Note. These figures are binned scatter plots of median per-pupil spending by county, absorbing county effects. The curve is a quadratic fit plot for the underlying data. Expenditures are adjusted to 2013 dollars. A county is designated as high-poverty if its median poverty rate for individuals aged 5-17 is over 25 percent between 1996-2013, and designated as low-poverty otherwise.
46
Figure 2. Associations between per-pupil spending and student math and reading performance (by county-year observation)
Note. Spending is measured here as log per-pupil current expenditures in 2013 inflation-adjusted dollars. County-level reading and math scores are measured in standard deviations, normalized by subject, grade and year. Each point represents a single county-year observation.
47
Figure 3a. Change in average county-level property revenues by years since property revaluation year
Figure 3b. Change in average county-level expenditures by years since property revaluation year
010
2030
40Pe
rcen
t cha
nge
in p
rope
rty re
venu
es
-4 -2 0 2 4Years since property revaluation
-10
010
20Pe
rcen
t cha
nge
in to
tal c
ount
y ex
pend
iture
s
-4 -2 0 2 4Years since property revaluation
48
Figure 4. County expenditures during Board of County Commissioners political party turnover events (Republican to Democrat)
Note. In this graph the blue line shows a lowess plot of per-pupil total county expenditures before and after a political party switch in the county board of commissioners. The gray lines are 95% confidence intervals.
Figure 5. Registered voter political affiliation during Board of County Commissioners political party turnover events (Republican to Democrat)
7600
7700
7800
7900
8000
8100
Per-
Pupi
l Cou
nty
Expe
nditu
res
-5 0 5Years Since Party Turnover
.348
.35
.352
.354
.356
.358
Perc
ent o
f Reg
iste
red
Vote
rs D
emoc
ratic
-5 0 5Years Since Party Turnover
49
Note. In this graph the blue line shows a lowess plot of the percent of county voters that are registered as Democrats before and after a political party switch in the county board of commissioners. The gray lines are 95% confidence intervals. Figure 6a. Graphed first stage estimates for total county expenditures
Figure 6b. Graphed first stage estimates for instructional salaries
-600
-400
-200
020
0Pe
r-pup
il co
unty
exp
endi
ture
s
0 2 4 6 8Years since revaluation
Republican Democratic
-40
-20
020
4060
Per-p
upil
inst
ruct
iona
l sal
arie
s
0 2 4 6 8Years since revaluation
Republican Democratic
50
Figure 6c. Graphed first stage estimates for non-instructional salaries
Note. For Figures 6a, 6b, and 6c, all predicted values are scaled as their difference in per-pupil expenditures from counties who are in a revaluation year with a Democratic majority board of commissioners.
-40
-20
020
Per-p
upil
non-
inst
ruct
iona
l sal
arie
s
0 2 4 6 8Years since revaluation
Republican Democratic
51
Figure 7a. Mediation model of math test performance: The role of non-academic behaviors in explaining expenditure effects
Figure 7b. Mediation model of reading test performance: The role of non-academic behaviors in explaining expenditure effects
Note. These models are estimated in a structural equation model using maximum likelihood with missing values (MLMV) estimation. Not pictured above, but controlled for in the model, are student demographic characteristics, prior year test scores, and grade and year effects.
52
Figure 8a. Non-parametric effects of predicted instructional expenditures on math performance
Note. The dark solid line represents kernel-weighted local polynomial smoothed regressions. The light solid lines represent 95% confidence intervals. The per-pupil spending variable has been estimated as a function of the vector of instrumental variables, and de-meaned by county and year.
-.20
.2.4
.6M
ath
perfo
rman
ce (S
Ds)
0 500 1000 1500Local instructional spending per-pupil
53
Figure 8b. Non-parametric effects of predicted non-instructional expenditures on math performance
Note. The dark solid line represents kernel-weighted local polynomial smoothed regressions. The light solid lines represent 95% confidence intervals. The per-pupil spending variable has been estimated as a function of the vector of instrumental variables, and demeaned by county and year.
-.10
.1.2
.3M
ath
perfo
rman
ce (S
Ds)
0 50 100 150 200Local non-instructional spending per-pupil
54
Figure 8c. Effect of predicted total county expenditures on math performance (Kernel-weighted local polynomial smoothing)
Note. The dark solid line represents kernel-weighted local polynomial smoothed regressions. The light solid lines represent 95% confidence intervals. The per-pupil spending variable has been estimated as a function of the vector of instrumental variables, and de-meaned by county and year.
-1-.5
0.5
Mat
h pe
rform
ance
(SD
s)
5000 10000 15000 20000 25000Total county spending per-pupil
55
Table 1. Descriptive statistics of student cross-sectional sample (2008-2009) Variable Mean SD Parental education Less than high school 0.051 (0.219) High school graduate 0.389 (0.487) Some college 0.154 (0.361) 4 year degree 0.297 (0.457) Graduate degree 0.109 (0.312) Race/ethnicity White 0.556 (0.497) Black 0.273 (0.446) Hispanic 0.098 (0.297) Other 0.072 (0.258) Behavioral indicators Absences 7.346 (7.849) Disciplinary infractions 0.397 (1.411) Other indicators Limited English Proficiency 0.062 (0.241) Exceptional status 0.112 (0.316) Eligible free/reduced price lunch 0.538 (0.499) Gifted status (ELA) Gifted status (Math)
0.171 0.180
(0.376) (0.384)
Note. Because not all of the student-level covariates above are included during all years of the observed student dataset, I estimate all descriptive statistics instead on a single cross-section of data (the 2008-2009 school year).
56
Table 2. First stage equation results: Effect of county-level instrumental variables on local per-pupil spending by purpose.
Per-pupil local salary expenditures
Non-Instruction Instruction Total county
spending
Years since revaluation 3.637** 14.368** 165.121** (0.070) (0.084) (1.013) Years since squared -0.494** -2.523** -24.087**
(0.010) (0.012) (0.144) Republican majority -17.563** 58.254** -447.528**
(0.170) (0.206) (2.472) Years X Rep majority 6.849** -17.552** -102.590** (0.100) (0.123) (1.457) Squared X Rep majority -0.796** 3.010** 12.771** (0.020) (0.018) (0.216) County fixed effects X X X Year fixed effects X X X F-Statistic 9,946 14,530 6,060 Observations 7,350,836 7,350,836 7,350,836 Number of counties 100 100 100 ** p<0.01, * p<0.05, + p<0.1 Robust standard errors in parentheses, clustered by county; covariates include: county population, political, and economic indicators, lagged math and reading scores (for grades 4-8), eighth grade math and reading scores (for grades 9-12), race/ethnicity indicators, gender indicators, and grade indicators.
57
Table 3. Test score IV estimates: Effect of instructional and non-instructional spending on end of grade and end of course achievement.
Grades 4-8:
Math Grades 4-8:
Reading Grades 9-12:
Sciences Grades 9-12: Humanities
$100 per-pupil local spending Non-instructional spending (prior 2 years)
0.067** 0.014+ 0.099** -0.001 (0.006) (0.007) (0.021) (0.002)
Instructional spending (prior 2 years)
0.016** 0.029** 0.078** 0.000 (0.002) (0.001) (0.002) (0.001)
Other county expenditures 0.000 -0.001* 0.000 -0.000 (0.000) (0.000) (0.000) (0.000) County fixed effects X X X X Year fixed effects X X X X Observations 7,766,298 7,332,072 2,319,431 2,527,446 Number of counties 100 100 100 100 ** p<0.01, * p<0.05, + p<0.1 Robust standard errors in parentheses, clustered by county; covariates include: county population, political, and economic indicators, lagged math and reading scores (for grades 4-8), eighth grade math and reading scores (for grades 9-12), race/ethnicity indicators, gender indicators, and grade indicators.
58
Table 4. Behavioral IV estimates: Effect of instructional and non-instructional spending on absences, infractions, and extra-curricular activities
Grades 4-8: Absences
Grades 4-8: Infraction
Grades 9-12: Activities
Per-pupil local spending Non-instructional spending (prior 2 years)
-0.577** -0.988** 0.139** (0.031) (0.061) (0.009)
Instructional spending (prior 2 years)
-0.029** -0.012 0.000 (0.001) (0.000) (0.000)
Other county expenditures 0.005** 0.000+ -0.000 (0.000) 0.000 (0.000 County fixed effects X X X Year fixed effects X X X Observations 13,107,075 2,322,119 3,412,120 Number of counties 100 100 100 ** p<0.01, * p<0.05, + p<0.1 Robust standard errors in parentheses, clustered by county; covariates include: county population, political, and economic indicators, lagged math and reading scores (for grades 4-8), eighth grade math and reading scores (for grades 9-12), race/ethnicity indicators, gender indicators, and grade indicators.
59
Table 5. Test score IV estimates by county poverty level: Effect of instructional and non-instructional spending on end of grade and end of course achievement.
Grades 4-8: Math
(Low pov.)
Grades 4-8: Reading
(Low pov.)
Grades 4-8: Math
(High pov.)
Grades 4-8: Reading
(High pov.) $100 per-pupil local spending Non-instructional spending (prior 2 years)
0.012** -0.002 0.239** 0.077* (0.002) (0.013) (0.085) (0.036)
Instructional spending (prior 2 years)
0.004** 0.006** 0.065 0.095* (0.001) (0.002) (0.043) (0.041)
Other county expenditures 0.000 -0.000+ 0.020** -0.000 (0.000) (0.000) (0.002) (0.001) County fixed effects X X X X Year fixed effects X X X X Observations 5,481,317 5,472,820 1,857,700 1,854,716 Number of counties 100 100 100 100 ** p<0.01, * p<0.05, + p<0.1 Robust standard errors in parentheses, clustered by county; covariates include: county population, political, and economic indicators, lagged math and reading scores (for grades 4-8), eighth grade math and reading scores (for grades 9-12), race/ethnicity indicators, gender indicators, and grade indicators.
60
APPENDIX
Appendix Figure A1. County variation in per-pupil expenditures from 1997 to 2013
Instructional total spending Non-instructional total spending
Legend: Total per-pupil expenditures. Expenditures are inflation-adjusted to 2013 dollars.
61
Appendix Table A1. Purpose codes and expenditure definitions
Purpose Label Definition
5100 Regular instructional programs
Instructional activities designed primarily to prepare pupils for activities as citizens, family members, and workers, as contrasted with programs to improve or overcome physical, mental, social, and/or emotional handicaps. Regular instructional programs include grades K-12.
5820 Attendance-social work services
Activities which are designed to improve pupil attendance at school and which attempt to prevent or solve pupil problems involving the home, the school, and the community.
5830 Guidance and psychological services
Activities of counseling pupils and parents, providing consultation with other staff members on learning problems, assisting pupils in personal and social development, assessing the abilities of pupils, assisting pupils as they make their own educational and career plans and choices, providing referral assistance, and working with other staff members in planning and conducting guidance programs for pupils
5840 Health services Physical and mental health services that are not direct instruction. Included are activities that provide pupils with appropriate medical, dental, and nursing services.
Source: NC DPI, 2007
62
Appendix Table A2. Illustration of county revenues instrumental variable for three fictional counties. County 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005
A 6 7 0 1 2 3 4 5 6 7 0 B 3 4 5 6 7 0 1 2 3 4 5 C 3 0 1 2 3 0 1 2 3 0 1
The values in each cell reflect the value that the primary Revaluation instrumental variable (years since revaluation) would take in each year and each county for these fictional counties. County A is shown to have revaluation years in 1997 and 2005; county B has a revaluation year in 2000; and county C has three revaluation years: 1996, 2000, and 2004. Most counties perform revaluations only at the required frequency of every eight years, but some choose to do so more frequently. All property revaluation years are set far in advance (typically over a decade).
Appendix Table A3. Comparison of OLS and IV estimates: Effect of instructional and non-instructional spending on end of grade achievement.
Grades 4-8: Math (IV)
Grades 4-8: Reading
(IV)
Grades 4-8: Math (OLS)
Grades 4-8: Reading (OLS)
$100 per-pupil local spending Non-instructional spending (prior 2 years)
0.067** 0.014+ 0.001** 0.004 (0.006) (0.007) (0.000) (0.036)
Instructional spending (prior 2 years)
0.016** 0.029** 0.014** 0.002 (0.002) (0.001) (0.002) (0.000)
County fixed effects X X X X Year fixed effects X X X X Observations 7,766,298 7,332,072 7,344,287 7,332,796 Number of counties 100 100 100 100 ** p<0.01, * p<0.05, + p<0.1 Robust standard errors in parentheses, clustered by county; covariates include: county population, political, and economic indicators, lagged math and reading scores (for grades 4-8), eighth grade math and reading scores (for grades 9-12), race/ethnicity indicators, gender indicators, and grade indicators.
63
Appendix Table A4. Robustness check: Effect of expenditures in year t+1 on outcomes in year t
Grades 4-8:
Math Grades 4-8:
Reading Grades 4-8: Absences
Grades 4-8: Infractions
$100 per-pupil local spending Non-instructional spending (prior 2 years)
0.001 -0.234+ 0.028 -0.008 (0.013) (0.142) 0.853 (0.029)
Instructional spending (prior 2 years)
0.003 0.015 0.013 0.001 (0.143) (0.013) (0.070) (0.003)
County fixed effects X X X X Year fixed effects X X X X Observations 7,344,287 7,332,796 13,107,075 2,322,119 Number of counties 100 100 100 100 ** p<0.01, * p<0.05, + p<0.1 Robust standard errors in parentheses, clustered by county; covariates include: county population, political, and economic indicators, lagged math and reading scores (for grades 4-8), eighth grade math and reading scores (for grades 9-12), race/ethnicity indicators, gender indicators, and grade indicators.